基于惩罚回归的高噪声流量分类  

High Noise Traffic Classification based on Penalty Regression

作　　者：白凯毅 盛志伟 黄源源 AI Kaiyi;SHENG Zhiwei;HUANG Yuanyuan(College of Cyberspace Security,Chengdu University of Information Technology,Chengdu 610225,China;Sichuan Provincial Key Laboratory of Advanced Cryptography Technology and System Security,Chengdu 610225,China)

机构地区：[1]成都信息工程大学网络空间安全学院,四川成都610225 [2]先进密码技术与系统安全四川省重点实验室,四川成都610225

出　　处：《成都信息工程大学学报》2025年第2期125-131,共7页Journal of Chengdu University of Information Technology

基　　金：国家重点研发计划资助项目(2022YFB3103103);四川省重点研发计划资助项目(2022YFS0571);四川网络文化研究中心资助项目(WLWH22-18);四川省自然科学基金资助项目(2022NSFSC0557)。

摘　　要：针对网络流量数据容易受到干扰的现实情况,引人带噪声标签学习的思想,并人为添加噪声以模糊化特征。先建立特征和标签之间的线性关系,然后用mean-shift参数识别噪声数据。通过人工添加对称噪声和非对称噪声模拟现实情况下的各种干扰信息。由此提出一个基于L2正则的高噪声流量分类模型(PR-2),通过将流量转换为图像并应用L2正则化方法来处理带噪声的标签,以提高高噪声流量下分类模型的性能。在USTC-TF2016数据集上验证了本方法的有效性,并与LSTM、BiTCN、BoAu、CL、INCV、FINE方法进行对比。实验结果表明,PR-2方法在对称噪声和非对称噪声的噪声比为0.8的情况下仍能取得95.16%和86.15%的准确率,证明其在处理高噪声数据方面的有效性和可用性。This paper aims to address the data quality issues faced in the field of high noise traffic classification.In re-sponse to the reality that network traffic data is prone to interference,the idea of noisy label learning(LNL)is introduced,and noise is artificially added to blur features.Firstly,establish a linear relationship between features and labels,and then use non-zero mean-shift parameters to identify noisy data.Simulate various interference information in real situations by manually adding symmetric and asymmetric noise.Therefore,this paper proposes a high-noise traffic classification model based on L2 regularization(PR-2),which converts traffic into images and applies the L2 regularization method to process noisy labels to improve the performance of the classification model under high-noise traffic.The effectiveness of this method was validated on the USTC-TF2016 dataset and compared with LSTM,BiTCN,BoAu,CL,INCV,and FINE methods.The ex-perimental results show that the PR-2 method can still achieve 95.16%and 86.15%accuracy even when the proportion of symmetric and asymmetric noise is 80%,demonstrating its effectiveness and usability in processing high-noise data.

关 键 词：高噪声流量 流量分类 深度学习 带噪声标签学习 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]